Detecting novel metaphor using selectional preference information
Hessel Haagsma and Johannes Bjerva
University of Groningen
The Netherlands
Abstract
costs an arm and a leg.’ and the Dutch ‘Dit lesboek
kost een rib uit het lijf.’ (lit. ‘This textbook costs a
rib from the body.’) to be mapped to the same meaning representation, which should not include references to any of the body parts used in the idiomatic
expressions.
In this paper, we focus on one area of figurative
language processing, namely the detection of novel
metaphor, and especially the utility of selectional
preference features for this task. By doing this, we
hope to answer a two-fold question: can selectional
preference information be used to successfully detect novel metaphor? And how do generalization
methods influence the effectiveness of selectional
preference information?
Recent work on metaphor processing often
employs selectional preference information.
We present a comparison of different approaches to the modelling of selectional preferences, based on various ways of generalizing over corpus frequencies. We evaluate on
the VU Amsterdam Metaphor corpus, a broad
corpus of metaphor. We find that using only
selectional preference information is enough
to outperform an all-metaphor baseline classification, but that generalization through prediction or clustering is not beneficial. A possible explanation for this lies in the nature of
the evaluation data, and lack of power of selectional preference information on its own
for non-novel metaphor detection. To better investigate the role of metaphor type in
metaphor detection, we suggest a resource
with annotation of novel metaphor should be
created.
1
2
2.1
Introduction
Within natural language processing (NLP), there has
been an increasing interest in the processing of figurative language in recent years. This increasing interest is exemplified by the (NA)ACL workshops on
metaphor in NLP, and shared tasks like SemEval2015, Task 11, about sentiment analysis of figurative
language in Twitter.
The main benefits of improved treatment of figurative language within NLP lie with high-level tasks
dependent on semantics, such as semantic parsing.
For example, in multilingual semantic parsing, one
would like to see both the English ‘This textbook
Background
Types of metaphor in language
The first part of any work on metaphor processing is
is to arrive at a definition for ‘metaphor’.
A good starting point for this is the MIP framework (Pragglejaz Group, 2007), which has this as
the most important criterion for metaphor: ‘a lexical unit is metaphorical if it has a more basic contemporary meaning in other contexts than in the current context’ (paraphrased). Where ‘more basic’ is
defined as more concrete, related to bodily action,
more precise, and historically older. This is a very
broad definition, causing many meanings to be classified as metaphorical.
Examples of this are found in the VU Amsterdam Metaphor corpus, which is annotated using the
MIPVU procedure (Steen et al., 2010), an extension of MIP. In the corpus, we find a wide range of
10
Proceedings of The Fourth Workshop on Metaphor in NLP, pages 10–17,
c
San Diego, CA, 17 June 2016. 2016
Association for Computational Linguistics
metaphor from highly conventionalized metaphor,
such as have in Example 1, to more clear metaphorical cases, like rolling in Example 2.
(1)
Do the Greeks have a word for it?
(2) [. . .] the hillsides ceased their upward
rolling and curved together [. . .]
Three general categories of meaning can be distinguished. The first type of contextual meaning is
literal meaning, which is the most basic meaning of
a word. This kind of meaning is generally the least
problematic, since the meaning can be deduced from
the lexicon. The second type of meaning is conventional metaphorical meaning, i.e. a non-basic meaning of a word, which is also in the lexicon or sense
inventory.
Novel metaphors form a third type of contextual
meaning, one which is non-basic, but is not accounted for by the lexicon either. These are the most
problematic, since their meaning cannot be deduced
by selecting the correct sense from a sense inventory.
Aligned with the distinction between conventional metaphor and novel metaphor is the distinction between word sense disambiguation and
metaphor detection. That is, word sense disambiguation can be expected to cover literal and
conventionalized metaphorical meanings, whereas
metaphor processing is necessary to deal with novel
metaphorical meanings.
We argue here that the most important application of metaphor processing approaches lies not with
conventional metaphors, but with novel metaphors.
This type of metaphor is the most problematic for
other NLP applications, since it requires a deduction
of meaning from outside the sense lexicon. Therefore, it requires a dedicated metaphor processing
system, whereas more conventional metaphor can,
in principle, be handled by existing WSD systems.
This is not to say that metaphor processing research should exclusively focus on novel metaphor,
since insights about metaphor and metaphoricity can
definitely be useful for improving the handling of
conventional metaphor within a word sense disambiguation framework. Nevertheless, the approach
proposed in this paper aims only to deal with novel
metaphor. In addition, separating the treatment of
novel and conventional metaphor avoids the possi11
ble pitfall of trying to solve two essentially different
problems using the same method.
2.2
Selectional preference violation as a feature
for metaphor detection
Because, by definition, we cannot use a pre-existing
sense or meaning definition for novel metaphor,
novel metaphor processing provides a new kind of
challenge. On the other hand, with the ultimate goal
of integrating figurative language processing in a semantic parsing framework, novel metaphor is the
most problematic for deducing meaning, and is thus
the most relevant and interesting.
Another important aspect of novel metaphors is
that they tend to be less frequent than conventional
metaphor (and literals). This has the advantage of
making one feature commonly used for metaphor
detection, selectional preference violation, more effective.
The idea of using selectional preference violation as an indicator of metaphoricity goes far back
(Wilks, 1975; Wilks, 1978), and has been widely
used in previous work on metaphor detection (Fass,
1991; Mason, 2004; Jia and Yu, 2008; Shutova et
al., 2010; Neuman et al., 2013; Wilks et al., 2013).
Using selectional preference violation as a heuristic works well, but has the fundamental shortcoming
that selectional preferences are based on frequency;
selectional preference data are obtained from a corpus by counting how often words occur together in a
certain relation. Metaphor, on the other hand, is defined by basicness of meaning, and not frequency of
meaning. Although these two are correlated, there
are also many cases where the figurative sense of a
word has become more frequent than its original, literal sense.
The result of this is that metaphor detection systems based on selectional preferences erroneously
classify low-frequent literals as metaphorical and
high-frequent metaphors as literal. However, if we
intend only to capture novel metaphors, the selectional preference violation approach is less vulnerable to these errors, since novel metaphors are less
likely to be mistakenly classified as literal due to
their lower frequency.
As such, we hypothesize that this feature, when
used correctly, can be very useful for metaphor detection. To this purpose, we examine new ways of
generalizing over selectional preferences obtained
from a corpus.
3
Methods
Selectional preference information can be automatically gathered from a large, parsed corpus. This allows for counting the occurrences of specific pairs,
such as ‘evoke-excitement’. Based on this, we can
then calculate conditional probabilities or a measure
like selectional association (Resnik, 1993).
However, even a very large corpus will not yield
full coverage of all possible pairs. This could cause
the system to flag, for example ‘evoke-exhilaration’
as a selectional preference violation and thus as
metaphoric, simply because this combination occurs
very infrequently in the corpus.
The solution to this problem is to do some sort of
generalization over these verb-noun pairs. One way
of doing this is to perform clustering the verbs and
nouns in semantically coherent classes. We test two
clustering methods: Brown clustering (Brown et al.,
1992), and a novel method, which relies on k-means
clustering of word embeddings.
In addition to these clustering methods, we also
look at using word embeddings directly in a classifier, in order to prevent the information loss inherent in a clustering approach. These approaches are
compared to performance based on using the selectional preference information without any generalization and a most-frequent class baseline.
3.1
Corpus frequencies
Before we can apply generalization methods over
selectional preferences, we need to gather selectional preference frequencies from a large corpus of
English. We use a recent dump (13-01-2016)1 of
the English part of Wikipedia. The corpus was preprocessed in order to extract only the raw text using
WikiExtractor2 , resulting in a raw text corpus of approximately 1.6 billion words.
We use Wikipedia because of its size, which
should help reduce sparsity for word pair frequencies, its range of topics, which increases the coverage in terms of word types. The downside of using
Wikipedia is that it contains only one genre, namely
1
2
https://dumps.wikimedia.org/enwiki/20160113/
http://medialab.di.unipi.it/wiki/Wikipedia Extractor
scientific or encyclopaedia-style text. In contrast,
the data used for evaluation is taken from different parts of the BNC, which represent different genres. Nevertheless, we assume that the larger size and
number of topics of the Wikipedia corpus compensate for this.
As a dependency parser, we use the spacy.io3
Python library, because of its speed, high accuracy
and ease-of-use. From the parsed corpus, all verbnoun pairs with the labels nsubj and dobj were
extracted and counted. This yielded 101 million
verb-subject triples of 14.0 million distinct types and
67 million verb-object triples of 7.8 million distinct
types. The triples contained 175,630 verb lemma
types and 1,982,512 noun lemma types.
3.2
Selectional preference information can be used in
three ways: as a conditional probability, a selectional preference strength for the verb, and as selectional association between a verb and a noun. The
conditional probability is defined as follows:
P (n, v)
c(n, v)
≈
(1)
P (v)
c(v)
Where P (n|v) is the probability of seeing noun n
occurring with verb v. This can be approximated by
using the counts c from the corpus.
For the other two metrics, we follow Shutova
(2010) in using the metrics proposed by Resnik
(1993), selectional preference strength (SPS) and
selectional association (SA). Selectional preference
strength is the information gain of a verb. This is an
indicator of how selective the verb is in the choice
of its arguments, and could thus be useful in filtering out weak-preference verbs.
P (n|v) =
SP S(v) =
X
n∈N
P (n|v) ∗ log
P (n|v)
P (n)
(2)
More directly useful is the selectional association
(SA) measure, which is an indication of how typical
a noun is in its occurrence with a certain verb. It is
defined as follows:
SA(v, n) =
3
12
Selectional preference metrics
1
P (n|v)
∗ P (n|v) ∗ log
(3)
SP S(v)
P (n)
https://spacy.io
3.3
Clustering
For the Brown clustering, we use the pre-trained
clusters from Derczynski et al. (2015). They provide clusterings4 with different numbers of clusters,
which makes it easy to find an optimal number of
clusters. We use the clusters trained on 64M words
of newswire text.
For k-means clustering of word embeddings, we
use the GloVe word embeddings (Pennington et al.,
2014), with 300 dimensions, trained on a 840B
word corpus and cluster the embeddings of the
top 400,000 most frequent words. Clustering is
done using the k-means clustering implementation
from scikit-learn (Pedregosa et al., 2011), with kmeans++ initialization, a maximum of 500 iterations, and the best clustering out of 10 runs with random initialization in terms of inertia.
3.4
Word embeddings-based probability
predictions
In the word embeddings-based learner setting, rather
than doing generalization through clustering, we
use the word embeddings for generalization directly.
This is done by modelling the conditional probabilities directly, based on the word embeddings of the
noun and verb involved. We expect this to benefit
generalization, since it avoids the information loss
inherent in clustering when dealing with distributed
representations – clustering (and thus discretizing)
such representations effectively removes any notion
of similarity between representations.
We use a neural network implemented in Theano
(Bergstra et al., 2010; Bastien et al., 2012) using
Keras. As input, we take the concatenation of the
distributed representations of each verb and its subject/object. The output of the network is a continuous variable, namely the log probability of the construction at hand. We use a single hidden layer containing 600 hidden units, with a sigmoid activation
function. Optimisation is done using the ADAM optimiser (Kingma and Ba, 2014). The loss function
used is the mean squared error. Dropout is applied
as regularisation after the hidden layer (p = 0.5)
(Srivastava et al., 2014). We use a batch size of
2,000 and train over the course of 200 epochs. The
resulting predictions are used as the Predicted Log4
Probability (P-LP) feature. Metaphor detection results using this feature are presented in Table 4.
3.5
We evaluate the approaches on the VU Amsterdam
Metaphor corpus (VUAMC)5 . The corpus is preprocessed by extracting all raw text, and marking each
word with the attribute function="mrw", which
indicates a metaphor-related word, as metaphor.
This results in a corpus of 40,622 verbs, of which
23,069 have at least one subject or object relation.
We split the data three-way: verbs with only a subject (13,466 pairs), verbs with only an object (3,913
pairs), and verbs with both a subject and an object
(5,539 triples).
We use this corpus since it is the largest corpus
with metaphor annotations available. The downside
is that it includes a large number of very conventionalized metaphors, and there is no annotation of novel
metaphor or a degree of metaphoricity.
Evaluation is performed using a logistic regression classifier, which takes one or more of the selectional preference-based features. The features
used are: conditional probability, log probability, selectional preference strength and selectional association, for subject-verb and object-verb pairs, and
based on corpus frequencies, predictions or clusterings. We use only those features which are available,
i.e. for the subject dataset we only use the four features for subject-verb pairs, while in the subject-andobject dataset we use eight features, four based on
the subject-verb pair and four based on the objectverb pair.
Model fitting and evaluation is done using 10-fold
cross validation, and l2-regularization with strength
1 is applied. In case of missing data points (e.g.
no cluster available for the verb/noun), the majority class (non-metaphor) is assigned. As a baseline,
we calculate the score when classifying all items as
metaphor.
In addition, we evaluate the effect of re-weighting
examples. Beigman Klebanov et al. (2015), who
also evaluated their metaphor detection system on
the VUAMC, showed that re-weighting training examples can have a large impact on performance, due
to the large class imbalance in the VUAMC (mostly
5
http://derczynski.com/sheffield/brown-tuning
13
Evaluation
http://www.vismet.org/metcor/documentation/home.html
Data
BL CP
LP P-LP SPS SA
Subject 23.0 0.0 0.0
Object 50.8 0.0 3.2
Both
53.4 0.0 18.1
Table 1:
0.0
1.4
0.7
All
Data
0.0 0.0 1.3
0.0 0.0 2.4
0.0 2.3 32.1
lectional preference information without any generalization,
Features:
all-metaphor BaseLine
(BL), Conditional Probability (CP), Log-Probability (LP), word
embeddings-based Predicted Log-Probability (P-LP), Selectional Preference Strength (SPS), Selectional Association (SA),
combination of CP, LP, SPS and SA (All).
LP P-LP SPS
SA
All
Table 2:
F1-scores on metaphor identification using se-
lectional preference information without any generalization,
with varying features.
Features:
all-metaphor BaseLine
(BL), Conditional Probability (CP), Log-Probability (LP), word
embeddings-based Predicted Log-Probability (P-LP), Selectional Preference Strength (SPS), Selectional Association (SA),
combination of CP, LP, SPS and SA (All). Re-weighting of
non-metaphor). They find that re-weighting examples increases F1-score by sacrificing precision for a
large increase in recall.
4
CP
Subject 23.0 24.5 24.5 23.2 20.9 26.4 33.6
Object 50.8 53.4 45.6 49.2 49.0 51.2 47.6
Both
53.4 54.2 44.3 50.0 50.5 53.8 57.8
F1-scores on metaphor identification using se-
with varying features.
BL
Results
In this section, we present the results of evaluating
each of the approaches to modelling selectional preference information on the VUAMC, as described
previously. Table 4 shows the results for metaphor
detection without any form of generalization. Table 4 shows the same, but with re-weighting of examples.
The most striking difference here is the difference between the two sets of results. Without reweighting, performance is poor. The classifier classifies almost everything as non-metaphorical for the
subject and object datasets, and for the subjectobject dataset, even the best feature set is still far
below the baseline.
In contrast, we see that performance with reweighting is a lot better for all datasets. For the
subject and subject-object datasets, the setting with
all features performs best, out-performing the baseline by 10.6% and 4.4% respectively. For the object
dataset, however, performance is highest with only
the most basic feature, the conditional probabilities,
out-performing the baseline by 2.6%. Clearly, the
re-weighting of of examples enables the classifier to
find a model that actually detects metaphor, rather
than defaulting to a majority-class baseline.
Tables 5 shows the results for Brown clustering,
Table 5 does the same for k-means clustering, and
the P-LP column in Table 4 shows the results for
word-embeddings based prediction. Since perfor14
training examples was applied.
mance using all features and re-weighting of examples worked best in the no-clustering setting, we
only report the results using these settings.
Results show that all generalization approaches
fail to improve over the non-generalization setting.
The Brown clustering seems to work better than
the k-means clustering, which in turn yields slightly
higher results than the predicted probabilities. The
differences between the datasets are not affected by
generalization, performance, relative to the baseline,
is highest for the subject-only data, whereas it is
lowest on the object-only data.
For both the Brown and k-means clustering, cluster size seems to only have a minor effect on performance, and no consistent pattern emerges from
the results. That is, there is no cluster size which
performs consistently better across data sets and/or
clustering methods.
5
Discussion
The most prominent conclusion we can draw from
the results is that, in the current set-up, generalization across selectional preference information using
clustering or prediction does not work. Although the
generalization approaches sometimes improve over
the baseline, they never out-perform the best results
from the no-generalization setting.
The idea behind generalization is that we sacrifice
some information (about specific words) for a gain
in robustness of the feature values, especially for
less frequent words and word combinations. Here,
clearly, the benefits do not outweigh the disadvantages. Since we explored a large range of cluster
Data BL
80 160 320 640 1280 2560 5120
Subj 23.0 26.3 28.8 27.9 25.9 26.3 26.6 25.3
Obj 50.8 48.7 47.7 45.3 46.9 44.7 44.6 46.2
Both 53.4 52.7 52.8 53.7 54.3 53.5 54.3 54.5
Table 3: F1-scores on metaphor identification using Brown
clustering for generalization, for varying numbers of clusters,
using all available features. Re-weighting of training examples
was applied.
Data BL
80 160 320 640 1280 2560 5120
Subj 23.0 24.2 23.5 30.7 28.6 24.4 23.6 22.9
Obj 50.8 40.4 44.8 45.8 44.2 48.9 48.8 49.8
Both 53.4 49.8 48.2 50.4 49.2 47.6 50.4 49.5
Table 4: F1-scores on metaphor identification using word
embedding-based k-means clustering for generalization, for
varying numbers of clusters, using all available features. Reweighting of training examples was applied.
sizes, which did not have a clear effect on detection performance, the root cause for this is likely to
be the semantic coherence of the clusters. That is,
either the clusters are not semantically coherent, or
they are not coherent in such a way that they form a
useful domain for metaphor detection.
As for the different datasets, we see a large performance difference in Table 4, where the model almost
always defaults to a majority class classification in
the object- and subject-only cases. Re-weighting
the training data, however, removes this effect completely. Even for the subject-only data, which has
the most skewed distribution (only 13% metaphor)
the re-weighting model works well.
Looking at instances classified as metaphor by
the best-performing system (57.8% F-score, Table 4)
reveals three main things about the system performance and the nature of the evaluation data.
First, all datasets, but especially the subject- and
object-only datasets contain a large proportion of
pronoun subjects and objects, for which selectional
preference information is not useful, unless pronoun
resolution is applied. Since the metaphor annotation
procedure does include resolving of referring expressions, this significantly hurts performance. An
obvious improvement therefore would be to com15
bine metaphor detection with coreference resolution.
Second, we find that a large part of the triples
classified as metaphor contain a light verb, such as
have, make, take, and put. Looking at the dataset as
a whole reveals that 85-90% of the triples containing
these verbs are classified as metaphor by the system.
In the VUAMC, these verbs are often annotated as
metaphor, since, due to their wide usage, they often
occur in a sense that is not the most basic. It could
be argued that these verbs, on the contrary, have such
an eroded meaning that they are never metaphorical
(Shutova and Teufel, 2010; Beigman Klebanov et
al., 2014).
Third, we notice that the amount of novel
metaphors in the VUAMC is minimal. Looking at
instances correctly classified by the best-performing
system yielded only one example of novel metaphor
(Example 3). Here, the verb escape is used in a
novel way; no matching sense is found in WordNet.
(3) [. . .] Adam might have escaped the file
memories for years,
We hypothesized that selectional preference information can be used successfully for detection of
novel metaphor, but the VUAMC contains all kinds
of metaphor, including a large number of highly
conventionalized metaphors. As such, even if the
hypothesis is true, we cannot see this in the current evaluation setting. Unfortunately, this presents
us from answering the research question posed in
the introduction; it is still unclear whether selectional preference information is useful for detection
of novel metaphor.
Due to the different kinds of metaphor in the corpus, the logistic regression model attempts to learn
a model for all kinds of metaphor. This is not possible, and as such almost all instances get classified
as the majority class, non-metaphorical. If novel
metaphor could be successfully detected with the
features used here, the regression learner might not
produce a model that achieves this, since there is a
large majority of conventional metaphor in the data.
In order to empirically evaluate how well these
features work for novel metaphor, a corpus of novel
metaphor, or a general corpus of metaphor with
metaphoricity annotations is required. Considering that this is not available, an interesting course
for future work would be to create such a resource,
and use that for the evaluation of systems for novel
metaphor detection.
This could be done as a completely new resource,
but also as an extension of existing resources’ annotation. The VUAMC for example, could be extended
by annotating whether metaphorically used words
are used in a novel sense, i.e. an out-of-vocabulary
sense. This would fit nicely with the MIPVU procedure, which is already strongly dependent on a fixed
sense lexicon. An alternative would be to annotate metaphoricity or ‘novelty of metaphor’ as scalar,
based on human judgements. This would also allow
resources to distinguish more easily between types
of metaphor.
Acknowledgements
We would like to thank the anonymous reviewers
for their insightful and constructive comments, and
Beata Beigman Klebanov for pointing out the utility
of re-weighting the training data.
References
Frédéric Bastien, Pascal Lamblin, Razvan Pascanu,
James Bergstra, Ian J. Goodfellow, Arnaud Bergeron, Nicolas Bouchard, and Yoshua Bengio. 2012.
Theano: new features and speed improvements. In
Deep Learning and Unsupervised Feature Learning
NIPS 2012 Workshop.
Beata Beigman Klebanov, Chee Wee Leong, Michael
Heilman, and Michael Flor. 2014. Different texts,
same metaphors: Unigrams and beyond. In Proceedings of the Second Workshop on Metaphor in NLP,
pages 11–17.
Beata Beigman Klebanov, Chee Wee Leong, and Michael
Flor. 2015. Supervised word-level metaphor detection: Experiments with concreteness and reweighting
of examples. In Proceedings of the Third Workshop on
Metaphor in NLP, pages 11–20.
James Bergstra, Olivier Breuleux, Frédéric Bastien, Pascal Lamblin, Razvan Pascanu, Guillaume Desjardins,
Joseph Turian, David Warde-Farley, and Yoshua Bengio. 2010. Theano: a CPU and GPU math expression
compiler. In Proceedings of the Python for Scientific
Computing Conference (SciPy), June.
Peter F. Brown, Peter V. deSouza, Robert L. Mercer, Vincent J. Della Pietra, and Jenifer C. Lai. 1992. Classbased n-gram models of natural language. Computational Linguistics, 18(4):467–479.
16
Leon Derczynski, Sean Chester, and Kenneth S. Bøgh.
2015. Tune your Brown clustering, please. In Proceedings of RANLP 2015, pages 126–133.
Dan Fass. 1991. met*: A method for discriminating
metonymy and metaphor by computer. Computational
Linguistics, 17(1):49–90.
Yuxiang Jia and Shiwen Yu. 2008. Unsupervised Chinese verb metaphor recognition based on selectional
preferences. In 22nd Pacific Asia Conference on Language, Information and Computation, pages 207–214.
Diederik Kingma and Jimmy Ba. 2014. Adam: A
method for stochastic optimization. arXiv preprint
arXiv:1412.6980.
Zachary J. Mason. 2004. CorMet: A computational,
corpus-based conventional metaphor extraction system. Computational Linguistics, 30(1):23–44.
Yair Neuman, Dan Assaf, Yohai Cohen, Mark Last,
Shlomo Argamon, Newton Howard, and Ophir
Frieder. 2013. Metaphor identification in large texts
corpora. PLOS ONE, 8(4).
F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel,
B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer,
R. Weiss, V. Dubourg, J. Vanderplas, A. Passos,
D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay. 2011. Scikit-learn: Machine learning in Python.
Journal of Machine Learning Research, 12:2825–
2830.
Jeffrey Pennington, Richard Socher, and Christopher D.
Manning. 2014. GloVe: Global vectors for word representation. In Proceedings of EMNLP 2014, pages
1532–1543.
Pragglejaz Group. 2007. MIP: A method for identifying
metaphorically used words in discourse. Discourse,
Metaphor and Symbol, 22(1):1–39.
Philip Stuart Resnik. 1993. Selection and Information:
A class-based approach to lexical relationships. Ph.D.
thesis, University of Pennsylvania.
Ekaterina Shutova and Simone Teufel. 2010. Metaphor
corpus annotated for source-target domain mappings.
In Proceedings of LREC 2010, pages 3255–3261.
Ekaterina Shutova, Lin Sun, and Anna Korhonen. 2010.
Metaphor identification using verb and noun clustering. In Proceedings of COLING 2010, pages 1002–
1010.
Ekaterina Shutova. 2010. Automatic metaphor interpretation as a paraphrasing task. In Proceedings of ACL
2010, pages 1029–1037.
Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky,
Ilya Sutskever, and Ruslan Salakhutdinov. 2014.
Dropout: A simple way to prevent neural networks
from overfitting. The Journal of Machine Learning
Research, 15(1):1929–1958.
Gerard J. Steen, Aletta G. Dorst, J. Berenike Herrmann,
Anna Kaal, Tina Krennmayr, and Trijntje Pasma.
2010. A method for linguistic Metaphor Identification:
from MIP to MIPVU. John Benjamins, Amsterdam.
Yorick Wilks, Lucian Galescu, James Allen, and Adam
Dalton. 2013. Automatic metaphor detection using large-scale lexical resources and conventional
metaphor extraction. In Proceedings of the First Workshop on Metaphor in NLP, pages 36–44.
Yorick Wilks. 1975. A preferential, pattern-seeking, semantics for natural language inference. Artificial Intelligence, 6(1):53–74.
Yorick Wilks. 1978. Making preferences more active.
Artificial Intelligence, 11(3):197–223.
17